Dear Reviewer AvnH,
Thank you for your feedback. We hope to address your questions and comments below.
> Q1:"The tests are mainly conducted on binary classification tasks, making it unclear if the findings can be generalized to other tasks."
1. Our primary motivation for using synthetic 2D classification datasets is their amenability to decision boundary visualization, which is central to our paper's focus as decision boundaries are typically studied in 2D space, and our proposed mechanism offers novel insights into the generalization abilities of in-context learning. NLP classification tasks, with their discrete token nature and variable input lengths, pose significant challenges for visualization of decision boundaries along x and y axes.
2. Our use of synthetic function datasets is firmly grounded in established practices for studying in-context learning. Several recent high-profile works have exclusively used synthetic data to investigate ICL/learning algos:
- NeurIPS 2022 paper (https://arxiv.org/pdf/2208.01066) focuses entirely on linear functions for ICL.
- ICLR 2023 paper (https://arxiv.org/pdf/2211.15661) explores learning algorithms for ICL and primarily conducts experiments on synthetic linear functions.
- Google's research (https://arxiv.org/abs/2303.03846) employs synthetic linear function classes (see Section 6).
- NeurIPS 2019 paper (https://arxiv.org/pdf/1905.11604) utilizes similar scikit-learn datasets to study decision boundaries for SDE algorithms (see Figure 1).
These papers demonstrate that our approach is quite prevalent in the field.
3. To extend beyond binary tasks, we conduct new experiments on 6 real-world NLP classification datasets. Note, to handle the high dimensionality of text embeddings, we projected them onto 2D space using t-sne. (Any dimensionality reduction technique will introduce confounders to the analysis, but this price is inevitable to extend our analysis). We experiment with a total of 6 widely-used NLP classification tasks, including both binary and multi-class settings. These include sentiment analysis and Textual Entailment Recognition, providing a broader perspective on the applicability of our approach. Our results, presented in **Figure 1 of the rebuttal PDF**, demonstrate similar non-smooth decision boundary characteristics in NLP tasks as observed in our synthetic datasets.
> Q2: "Although different model sizes are covered, these models are from different series. The reason for using these models should be clarified in the paper."
We use these LLMs to provide a comprehensive analysis across different sizes and architectures, representing current SoTA open-source LLMs. For the sizes, due to computational constraints in an academic setting and the expensive nature of querying decision boundaries with large grid points, we limited our analysis to models no larger than 13B parameters. Therefore to complete the size series and gain insights into smaller-scale models, we included the pruned llama 1.3B model.
> Q3: "The decision boundaries appear to be heavily influenced by the prompts, such as the design of labels and the order of in-context learning examples. ... explore the robustness of their results to variations in prompts, like different formats or synonym replacements?"
To address this concern, we have conducted experiments to explore how decision boundaries are affected by prompt formats, using 4 different synonyms for key terms. As shown in **Figure 2 of the rebuttal PDF**, we find that the LLM's decision boundary is indeed affected by the prompt format, which aligns with the importance of prompt engineering in ICL. However, the overall non-smoothness level of the decision boundaries remained consistent. We view prompt formats as another factor influencing the decision boundary rather than undermining our results, since our central observations regarding the non-smoothness of decision boundaries show similar patterns across different prompts. We will add this as another influencing factor in our paper.
> Q4: “Large Language Models are trained on vast corpora of text using unsupervised learning.” Shouldn't this be self-supervised learning?
Sorry for the typo! We will revise this.
> Q5: “For the open-source models, we use the approach described in 3.2 to get predictions. For the closed-source models, we use the next token generation as the prediction…” This is confusing since section 3.2 only mentions visualization methods. What are the differences in prompt design or experiments between these two approaches?
We apologize for the confusion. For both open and closed source models, we used the same prompts. For open source models, we used logits to get predictions for the class labels. For closed source models, we looked at the predictions since the logits were not available to us.
We hope this addresses your concerns. Please let us know if you have any questions.